Debug dim scatter

[MARD1NO]

相关细节

1. 针对后向传播，增加了scatter_update的scalar版本（由于torch的scatter_add没有scalar版本，故没有增加）
2. 对齐torch反向传播要求，即index.shape == source.shape才允许反传
3. 修改了单测index的生成逻辑，保证index不存在重复元素。原因是经过torch下测试，当存在重复元素时，GPU版的scatter_update出来的结果不一致（由于线程单独执行逻辑），而CPU版本由于其顺序执行的方式，使其输出结果一致。

新增接口

• oneflow.scatter 的前后向
• oneflow.scatter_add 的前后向
• oneflow.F.dim_scatter 的前后向
• oneflow.F.dim_scatter_add 的前后向
• oneflow.F.dim_scatter_scalar 的前后向
• oneflow.F.dim_scatter_add_scalar （只有前向，没有后向，这个 PyTorch 没有）

以上没有后向的算子，是因为 PyTorch 也没有后向。

文档生成：

[github-actions]

CI failed, removing label automerge

[github-actions]

CI failed, removing label automerge

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 140.2ms (= 7009.5ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 126.6ms (= 6329.3ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.11 (= 140.2ms / 126.6ms)

PyTorch resnet50 time: 83.5ms (= 4173.3ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 75.2ms (= 3758.0ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.11 (= 83.5ms / 75.2ms)

PyTorch resnet50 time: 57.4ms (= 2868.3ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 47.6ms (= 2381.0ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 1.20 (= 57.4ms / 47.6ms)

PyTorch resnet50 time: 48.9ms (= 2443.9ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 43.2ms (= 2160.2ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 1.13 (= 48.9ms / 43.2ms)

PyTorch resnet50 time: 41.4ms (= 2072.3ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 42.3ms (= 2114.5ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 0.98 (= 41.4ms / 42.3ms)